Model and credential providers
Operator connects Attack Harness to its model provider and authenticates requests on the host. The Harness receives the permitted model interface, not provider credentials. Interceptor separately configures model access for the target agent.
Choose a model provider
Set provider, codec_id, and model in a private JSON model profile. The codec
selects the native conversation and tool-call format. Use a model that supports
the selected API and tool use.
provider |
codec_id |
endpoint or routing fields |
|---|---|---|
openai-chat |
openai-chat-text-tools-v1 |
https://api.openai.com/v1/chat/completions |
openai-responses |
openai-responses-text-tools-v1 |
https://api.openai.com/v1/responses |
anthropic-messages |
anthropic-messages-text-tools-v1 |
https://api.anthropic.com/v1/messages; also set anthropic_version to the provider's required API header value in YYYY-MM-DD form. |
gemini-api |
gemini-text-tools-v1 |
https://generativelanguage.googleapis.com/v1beta/models/YOUR_MODEL:generateContent |
vertex-gemini |
gemini-text-tools-v1 |
https://YOUR_LOCATION-aiplatform.googleapis.com/v1/projects/YOUR_PROJECT/locations/YOUR_LOCATION/publishers/google/models/YOUR_MODEL:generateContent, or the applicable express-mode endpoint. |
azure-openai |
Chat or Responses codec above | Your resource's complete operation URL, such as https://YOUR_RESOURCE.openai.azure.com/openai/v1/chat/completions or /openai/v1/responses. |
litellm |
Chat or Responses codec above | Your gateway's complete HTTPS operation URL ending in /chat/completions or /responses. |
bedrock-converse |
bedrock-converse-text-tools-v1 |
Set region; omit endpoint. Operator uses the regional AWS endpoint. |
Replace placeholders with the endpoint and model or deployment identifier available
to your account. HTTP profiles take the complete operation URL, not a base URL.
Gemini URLs must contain the same model as model. Only Azure permits a query
string, with exactly one api-version parameter when required by the chosen route.
Do not put keys in URLs. Generation is non-streaming and Operator does not retry
model requests automatically.
Configure the model profile
Save the following as a private host file, for example /absolute/private/model.json,
and replace YOUR_MODEL. The token and response bounds are illustrative; choose
allowances supported by the model and your campaign budget.
{
"api_version": "operator.dev/model-provider-profile/v1alpha1",
"id": "campaign-model",
"provider": "openai-chat",
"codec_id": "openai-chat-text-tools-v1",
"model": "YOUR_MODEL",
"endpoint": "https://api.openai.com/v1/chat/completions",
"authentication": "secret-store",
"credential_id": "model-key",
"maximum_prompt_tokens": 32768,
"maximum_completion_tokens": 4096,
"maximum_response_bytes": 1048576,
"codec_options": {"instruction_role": "developer"}
}
Merge these fields into the installed Operator configuration, using actual absolute paths. Preserve its existing image, contract, target, limits, and Docker settings:
model:
profile_file: /absolute/private/model.json
credentials:
file: /absolute/private/credentials.yaml
maximum_prompt_tokens includes message and tool framing and is used to reserve
campaign tokens. Both token bounds must be positive; maximum_response_bytes must
be positive and at most 4 MiB. Oversized reservations can prevent work within the
campaign's remaining budget. Optional context_warning_percent is 1–99, default 80.
For a private TLS service, ca_certificates_pem can add trusted PEM certificates
without disabling certificate or hostname checks.
| Codec | Optional codec_options |
|---|---|
| Chat | instruction_role: developer (default) or system. |
| Responses | Native reasoning object. |
| Anthropic | Native thinking and output_config objects, including output_config.effort. |
| Gemini | Native thinking_config object. |
| Bedrock | Use an empty object. |
Choose options supported by the selected model. These settings are private to Operator; the Harness cannot choose authentication, endpoints, or credentials.
Select model authentication
Set authentication in the model profile. Secret-store login is a separate choice
in the credential configuration described below.
authentication |
Model providers | Additional profile fields |
|---|---|---|
secret-store or api-key |
All HTTP providers above | credential_id; resolve the model API key from a configured store. |
api-key-env |
All HTTP providers above | api_key_env naming a host variable; omit credential_id. |
workload-token |
Anthropic | credential_id for a stored bearer token; the administrator manages its issuance and renewal. |
workload-identity |
Bedrock, Azure OpenAI, Vertex | No key fields; use the host's configured cloud workload identity. Bedrock also needs region. |
aws-profile |
Bedrock | region and aws_profile; omit key fields. |
aws-environment |
Bedrock | region; the host supplies the AWS access-key pair and optional session token. |
google-adc |
Vertex | No key fields; explicitly use Google Application Default Credentials. |
azure-cli |
Azure OpenAI | No key fields; use the existing Azure CLI login. |
azure-client-secret |
Azure OpenAI | No key fields; use the host's Azure service-principal variables. |
none |
LiteLLM | Omit credential fields; use only for an intentionally unauthenticated gateway. |
For environment-key authentication, replace the example's authentication and
credential_id with the following fragment. Provision the variable's value through
your host's credential mechanism; the JSON contains only its name:
{"authentication": "api-key-env", "api_key_env": "CAMPAIGN_MODEL_KEY"}
For Bedrock, a complete model-profile example is:
{
"api_version": "operator.dev/model-provider-profile/v1alpha1",
"id": "campaign-bedrock",
"provider": "bedrock-converse",
"codec_id": "bedrock-converse-text-tools-v1",
"model": "YOUR_BEDROCK_MODEL_OR_INFERENCE_PROFILE",
"region": "us-east-1",
"authentication": "aws-profile",
"aws_profile": "YOUR_AWS_PROFILE",
"maximum_prompt_tokens": 32768,
"maximum_completion_tokens": 4096,
"maximum_response_bytes": 1048576,
"codec_options": {}
}
See host identity setup for login prerequisites.
Configure credential providers
Operator supports AWS Secrets Manager, Azure Key Vault, Google Cloud Secret Manager, HashiCorp Vault KV v2, and 1Password. A profile selects a store and its permitted secret locations; a credential reference gives a stored value a stable ID used by the model profile. The model provider and secret store can be in different clouds.
Create /absolute/private/credentials.yaml with the following structure. This
example reads a plain-text API key from AWS Secrets Manager through an explicit
host AWS profile. Replace the region, profile name, and secret location:
profiles:
- api_version: operator.dev/secret-store-profile/v1alpha1
kind: SecretStoreProfile
id: model-secrets
backend_kind: aws-secrets-manager
authentication: aws-profile
aws_profile: YOUR_AWS_PROFILE
region: us-east-1
allowed_locator_prefixes: [aws://us-east-1/simulation/model-key]
timeout_seconds: 10
max_value_bytes: 65536
max_cache_ttl_seconds: 300
credentials:
- api_version: operator.dev/host-credential-ref/v1alpha1
kind: HostCredentialRef
credential_id: model-key
store_profile_id: model-secrets
locator:
backend_kind: aws-secrets-manager
region: us-east-1
secret_id: simulation/model-key
version_stage: AWSCURRENT
value: {format: utf8}
cache_ttl_seconds: 30
This file contains references, never the API key itself. Protect it and the model
profile with mode 0600; use regular files owned by the execution account or root,
without symlinks. Operator accepts a single YAML document or JSON with the same
structure, at most 1 MiB; YAML aliases and anchors are not supported.
Backend fields and secret locations
To select another backend, retain the common schema, kind, IDs, limits, and value
selection above. Replace backend_kind in both the profile and locator, remove
fields belonging to the previous backend, and supply these fields:
| Backend | Profile fields | locator fields |
Example allowed prefix |
|---|---|---|---|
aws-secrets-manager |
region; authentication: workload-identity, aws-profile, or aws-environment; aws_profile only for named-profile access. |
Matching region, secret_id; optional version_id or version_stage. |
aws://us-east-1/simulation/ |
azure-key-vault |
HTTPS vault_url; authentication: workload-identity, azure-cli, or azure-client-secret. |
Matching vault_url, secret_name; optional version. |
https://YOUR_VAULT.vault.azure.net/secrets/simulation- |
gcp-secret-manager |
authentication: workload-identity or google-adc. |
project, secret_name, version such as latest. |
gcp://projects/YOUR_PROJECT/secrets/simulation- |
hashicorp-vault |
HTTPS vault_url; exactly one of vault_token_file or vault_proxy: true; optional vault_namespace, vault_ca_certificate. |
mount, path; optional version as a positive integer encoded as a string. |
vault://secret/simulation/ |
onepassword |
authentication: onepassword-desktop with onepassword_account, or onepassword-service-account without the account field. |
secret_reference: op://vault/item/[section/]field. |
op://simulation/ |
Cloud-store authentication defaults to workload identity, except that an explicit
aws_profile selects that profile when authentication is omitted. The locator
must match the backend and the configured AWS region or Azure vault. Allowlist
prefixes restrict selection; they do not grant access at the remote store.
For example, these are the replacement profile fields and locator for Google Secret Manager, merged into the same two records above:
# Fields in profiles[0]; remove region and aws_profile.
backend_kind: gcp-secret-manager
authentication: google-adc
allowed_locator_prefixes: [gcp://projects/YOUR_PROJECT/secrets/simulation-]
# Replace credentials[0].locator.
backend_kind: gcp-secret-manager
project: YOUR_PROJECT
secret_name: simulation-model-key
version: latest
Values, caching, and store access
Use value: {format: utf8} when the selected secret is the key itself. For a JSON
secret, select one top-level string with
value: {format: json-string-field, field: api_key}. Vault KV v2 requires this
explicit field selection. Binary values, empty keys, NUL, and line breaks in the
selected credential are rejected. The resolver reads secrets; it does not create,
rotate, list, or delete them.
timeout_seconds defaults to 10 and permits 1–60. max_value_bytes defaults to
65536 and permits at most 1048576. max_cache_ttl_seconds defaults to 300 and permits
up to 3600. A reference's cache_ttl_seconds defaults to zero (no caching) and cannot
exceed its profile's ceiling. A failed refresh does not fall back to a stale value.
Grant the chosen host identity only the access it needs:
| Store | Access and setup |
|---|---|
| AWS Secrets Manager | secretsmanager:GetSecretValue on the secret; customer-managed encryption keys may also require kms:Decrypt and an allowing key policy. |
| Azure Key Vault | Key Vault Secrets User, or secret Get under legacy access policies. Key Vault Reader alone cannot read values. |
| Google Secret Manager | Secret Manager Secret Accessor on the selected secret, an enabled Secret Manager API, and an enabled secret version. |
| Vault KV v2 | Read access to MOUNT/data/PATH through an already authenticated Agent or Proxy. |
| 1Password | Read access to the selected vault/item for the desktop account or service account. |
For Vault, vault_token_file names an absolute private token-sink file; it is reread
on uncached resolution. Alternatively, use an authenticating Proxy with
vault_proxy: true. The product does not perform Vault login. Keep TLS verification
enabled; vault_ca_certificate supplies a trusted certificate file when needed.
For 1Password desktop access, run as the signed-in desktop user, enable
Settings → Developer → Integrate with 1Password SDKs, and set
onepassword_account to the account name or UUID. For automation, select
onepassword-service-account and supply OP_SERVICE_ACCOUNT_TOKEN only to the host
process. No op CLI is required. References select one field; names or IDs are
accepted, and allowed prefixes must end on a component boundary with /.
Configure host identities
Run authentication setup as the account that will execute Operator. Model access and permission to read a secret are separate grants.
| Selection | Host prerequisites |
|---|---|
AWS workload-identity |
Web-identity token file and role, container role, or instance role. Shared AWS files are excluded. |
AWS aws-profile |
Explicit profile name; for SSO, run aws sso login --profile YOUR_AWS_PROFILE. |
AWS aws-environment |
AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY, and optionally AWS_SESSION_TOKEN in the host process. Shared profiles are ignored. |
Azure workload-identity |
Federated token file with AZURE_CLIENT_ID and AZURE_TENANT_ID, or managed identity. |
Azure azure-cli |
Existing az login session and az available on the process PATH. |
Azure azure-client-secret |
Host AZURE_TENANT_ID, AZURE_CLIENT_ID, and AZURE_CLIENT_SECRET. |
Google workload-identity |
Attached identity or external-account federation, including a configured GOOGLE_APPLICATION_CREDENTIALS file. |
Google google-adc |
Explicit Application Default Credentials; local setup uses gcloud auth application-default login, separately from gcloud auth login. |
AWS workload and named-profile modes reject static AWS credential environment
variables; choose aws-environment when using an exported access-key session.
For direct model access, grant Bedrock InvokeModel, Azure Cognitive Services
OpenAI User, or Vertex AI User/inference permission for the selected resource.
Google quota-project use can also require Service Usage Consumer.
In CI, AWS credential actions that export temporary keys use aws-environment;
Azure CLI login uses azure-cli; Google's federated credential file uses
workload-identity or google-adc. Keep the authenticated environment and files
available for the process lifetime. A detached service needs its own configured
account and environment and does not automatically inherit a terminal login.
Check the configuration
After saving private files and wiring them into the installed configuration, run:
operatorctl config check --config /absolute/private/config.yaml
operatorctl doctor --offline --config /absolute/private/config.yaml
Configuration checks do not retrieve secrets or prove model access. Inspect diagnostics for profile, identity, and permission errors, then use your chosen campaign to verify model access. Secret resolution during a campaign is audited without recording the value or locator. For failures, check the selected host account, credential ID, allowlist, store permissions, and model permissions.
Continue with campaign preparation. Upstream references provide additional detail on model profiles and credential stores; choose the source tag matching your installed release.